![\"humanizer](\"/api/label/examples/humanizer/asset/screenshot.png\")
A script and a skill for taking the machine sound out of a draft.
\n![\"5](\"/api/label/examples/humanizer/asset/docs/flags.svg\")
The script does the part that is mechanical: swapping utilize for use, deleting In today's fast-paced world,, dropping Additionally, from the head of a sentence. The skill does the part that needs a person: negative parallelism, em dash habits, sentences that end by restating themselves as significance.
Nothing here calls a model or the network. Every rule is a string rule you can read in scripts/humanize.mjs and delete if you disagree with it.
No dependencies. Node 18 or newer.
\ngit clone <this repo> && cd humanizer\nnode scripts/humanize.mjs --help\nOr link the binary:
\nnpm link\nhumanize draft.md\nReport, without changing the file. The rewritten text goes to stdout; the findings go to stderr, so you can pipe one and read the other.
\nnode scripts/humanize.mjs draft.md > clean.md\nRewrite in place.
\nnode scripts/humanize.mjs draft.md --fix\nFrom a pipe.
\npbpaste | node scripts/humanize.mjs - | pbcopy\nAs a check in CI. --quiet suppresses the report and exits 1 if anything was found, so it fails a build on a generated-sounding README.
node scripts/humanize.mjs README.md --quiet || echo "README needs an edit"\n| Code | Meaning |
|---|---|
0 | nothing rewritten, nothing flagged |
1 | something was rewritten or flagged |
2 | bad usage |
Applied automatically by --fix. These are substitutions that are almost always right.
| Category | Examples |
|---|---|
| Lexicon | delve into → look at, leverage → use, utilize → use, in order to → to, a myriad of → many, prior to → before |
| Openers | Additionally, Moreover, Furthermore, Notably, Ultimately, at the head of a sentence, deleted |
| Boilerplate | In today's fast-paced world,, In the ever-evolving landscape of X,, When it comes to |
| Promotional adjectives | seamless, robust, cutting-edge, game-changing, crucial, comprehensive, deleted before the noun |
| Hedged framing | It is important to note that, It should be noted that, deleted |
Capitalisation follows what was replaced, and the pass tidies the double spaces and stranded punctuation its own deletions create. It is idempotent: running it twice changes nothing the second time, which is what makes it safe in a pre-commit hook.
\nFenced code blocks, inline code spans and indented code are never touched. They are masked out before any rule runs and restored byte-for-byte afterwards, so `utilize()` survives a pass that rewrites the prose around it.
Reported with a line number. The fix depends on what you meant, so the script refuses to guess.
\n| Flag | Example | The repair |
|---|---|---|
negative-parallelism | It's not just a parser, it's a compiler. | Delete the negative half. Assert the positive. |
em-dash | the result — a fast one — shipped | Ask whether a comma, a colon or a full stop does the work. |
superficial-ing | ..., highlighting the importance of testing. | Cut the clause; it restates the sentence as significance. |
vague-attribution | Experts say the approach is sound. | Name the source or delete the claim. |
rule-of-three | fast, cheap, and reliable | Keep the two that carry information. |
skills/humanizer.md is the guidance for working through those, written to be dropped into an agent's skills directory.
Copy skills/humanizer.md into wherever your agent reads skills from. It tells the agent to run the script first, then work the flags by hand, and it carries the rules of thumb a regex cannot check: cut the summary paragraph, one claim per sentence, prefer the concrete number, keep the writer's jokes.
--diff mode that prints a unified diff instead of the file--quiet already exits non-zeroDetails of each rule live in the skill.
\nnpm test\nThree tests, using the Node test runner, no dependencies: that the rewrite pass fires and leaves code spans alone, that the flags catch what they claim to, and that a second pass is a no-op.
\nWe leverage the leverage of the lever becomes We use the use of the lever; there is no parser deciding which leverage is a noun.rule-of-three flag has false positives on genuine three-item lists. It is a flag, not an error.MIT. See LICENSE.